AMD and Intel have agreed to maintain a single library for AI workloads on CPUs. AMD will stop developing its own ZenDNN fork and move its Zen optimizations straight into oneDNN, the library Intel originally started.

For developers, this means one PyTorch build can pick the fastest code path on any x86 processor by itself. AMD has also joined the steering board of the foundation that oversees the project.

Why there were two near-identical projects

oneDNN holds hand-tuned building blocks for deep learning: convolutions, matrix multiplication and activation functions. PyTorch, TensorFlow, ONNX Runtime and OpenVINO rely on it whenever a model runs on a CPU. Historically the library was tuned for Intel chips.

AMD answered with ZenDNN, which kept the same API but targeted Zen and EPYC server chips. It ran well on AMD hardware, but teams with mixed fleets had to keep track of which version of the library sat inside their stack.

What changes

Zen optimizations will land in the main oneDNN branch, and the library will detect the CPU at runtime. Stock PyTorch on Ryzen or EPYC will make use of the architecture without extra plugins. No date has been given for the first code merges yet.

The companies are also working on AI Compute Extensions (ACE), a new set of x86 instructions for matrix math. PyTorch support is planned through oneDNN, and the GCC 17 compiler already understands them.

What it does not cover

The deal only covers CPUs. For its graphics cards and Instinct accelerators AMD keeps developing the ROCm stack. oneDNN itself is formally no longer an Intel project: it lives in the UXL Foundation under the Linux Foundation umbrella.

Our take

Good news for anyone running small models on ordinary servers without a GPU: less fiddling with hardware-specific builds and fewer surprises when moving between Intel and AMD. For now it is a plan, so expect the speedup in future PyTorch releases rather than tomorrow. Details in the 3DNews report.